Breaking the Training Barrier of Billion-Parameter Universal Machine Learning Interatomic Potentials

A team of researchers has shattered a major bottleneck in AI-for-Science by developing a system that trains billion-parameter physics models orders of magnitude faster. They introduced two key innovations: MatRIS-MoE, a massive Mixture-of-Experts model designed as a universal Machine Learning Interatomic Potential (uMLIP), and Janus, a novel high-dimensional distributed training framework specifically engineered for these complex models. uMLIPs are foundational AI models pre-trained on diverse datasets to perform quantum-accurate simulations of materials and molecules, but their training has been prohibitively slow due to the need for second-order derivatives and explosive computational overhead at scale.

Janus tackles this with hardware-aware optimizations, enabling it to run efficiently across two Exascale supercomputers. The system achieved a staggering peak performance of 1.2 EFLOPS in single precision, representing 24-35.5% of the theoretical peak, while maintaining over 90% parallel efficiency. This performance leap compresses the training timeline for a billion-parameter uMLIP from what previously took weeks down to mere hours. The work establishes a new benchmark for AI4S foundation models and provides the essential high-performance computing infrastructure to accelerate discoveries in chemistry, materials science, and drug design by making rapid, large-scale physical simulations feasible.